Under The Hood: AI meets public administration

20/05/2026

Last week, Barbara Allen and I published a piece in The Conversation on Aotearoa New Zealand’s Public Service AI Framework (Te Kawa & Allen, 2026). We suggested, perhaps a little provocatively, that it exemplifies a kind of “Pollyanna policy”: a governance instrument that names all the right principles but issues them without the armour that would convert aspirations into implementable and enforceable obligations. If my inbox is anything to go by, the argument seems to have resonated. On reflection, that argument was unfinished. It was more provocation than precision, and those are not quite the same thing.

This post introduces a new series on He Poneketanga that picks up where that piece left off. The series is called Under the Hood because it seeks to look beneath the surface of AI governance frameworks and principles to examine the machine beneath: the layered, contested, path-dependent landscape that characterises public administration in Aotearoa and, indeed, in most comparable jurisdictions.

“Under the hood” is how practitioners and technologists alike describe what a system actually does, as distinct from what it presents. It is also, as it happens, how one Australian public servant recently described the accountability problem at the heart of algorithmic governance: that public servants “don’t understand what is under the hood” of the systems they are being asked to govern (Rahman et al., 2026). The phrase captures both sides of what this series intends to examine: the conditions in which algorithmic systems are deployed and the algorithmic systems themselves, and the ethical infrastructure needed to hold both.

Why this series, and why now

There is no shortage of writing on AI and government. A great deal of it, however, falls into one of two categories.

The first is technical guidance literature: frameworks, maturity models, capability assessments, and readiness tools. These are often useful in their own right, but they tend to treat the public sector as a relatively undifferentiated surface onto which new technology is applied. If we are honest, they are mostly technocratic pieces that ignore the messy realities of social license.

The second is normative commentary in the AI ethics tradition: pieces that articulate principles without sustained attention to the practical reasons why such principles reliably fail to travel from the page to practice. If we are honest about these pieces, they are mostly abstract and difficult to apply, and context-neutral.

What is considerably thinner on the ground is work that takes the institutions themselves as the primary unit of analysis. Some recent scholarship is beginning to attend to the relational and organisational dynamics of AI adoption. Rahman et al. (2026), drawing on interviews with 37 public servants across 22 Australian government agencies, argue that generative AI adoption fundamentally involves relational processes: the co-creation of meanings that emerges from the interactions between users, technologies, and organisational contexts. Their account is valuable, and their emphasis on sensemaking, institutional norms, and legitimacy resonates with the concerns that animate this series. But such work has not yet been connected to the specific institutional conditions of public administration in Aotearoa, nor to Te Tiriti-grounded analysis and opportunity that the data sovereignty discourse offers. That is where this series aims to contribute.

The institutional question matters because public sector governance is rarely the clean, ordered structure that most technology writing tends to presuppose. It is, more typically, an accumulation of layer upon layer of policy, operational procedures, ministerial expectations, legislative obligations, and professional conventions. New instruments seldom replace older ones. They are added alongside them, often interacting in ways that are difficult to anticipate and harder still to manage.

A general-purpose technology such as AI is now being asked to operate within precisely this environment. Evidence of its effect is emerging not only from academic commentary but also from audit offices.

In May 2026, the Auditor General of Ontario released a performance audit of AI use in the Ontario government (Office of the Auditor General of Ontario, 2026). Ontario introduced an AI Strategy in November 2024 and developed an AI Framework built on six guiding principles, including that AI use should be transparent, accountable, non-discriminatory, and aimed at benefiting the people of Ontario. What the auditors found, however, was that the state lacked consistently effective processes and procedures to implement those principles. Roughly 60 per cent of the AI websites accessed by Ontario public servants were rated unsafe or unsecured. The one approved, secure generative AI platform accounted for just 6 per cent of staff usage; unapproved platforms accounted for 94 per cent. Only 3 per cent of staff had completed the Ministry’s AI training, which was not mandatory. The strategy itself lacked several key components when benchmarked against comparable jurisdictions, including detailed initiatives, milestones, measurable outcomes, and ethical benchmarking, including the identification of prohibited AI practices.

Ontario is not Aotearoa. But the pattern it illustrates is the same one Barb and I identified in our article: principles present, implementation still emergent. The distance between the assumptions embedded in current governance frameworks and institutional readiness is, on the available evidence, more substantial than is often acknowledged.

This series attempts to examine and describe that gap.

The argument in outline

Under the Hood will proceed from foundations to diagnosis, comparison, and construction. Each instalment will take a discrete institutional problem that algorithmic governance either creates or exposes, ground it in the emerging academic literature, and connect it to the specific conditions of public administration in Aotearoa. The territory I intend to cover is set out below, though it may shift as the literature develops and as readers raise questions I had not anticipated.

When “AI” is too narrow a frame. The governance problems this series addresses are not unique to machine learning or generative systems. They are features of algorithmic decision-making more broadly. The Post Office Horizon scandal involved a database, not a neural network. Robodebt was an automated income-averaging calculation. INCIS and Novopay would not pass a narrow definitional test for “AI” either. This first post asks what happens when we define the problem too narrowly and quarantine precisely the lessons most relevant to the present moment.

Te Tiriti, tino rangatiratanga, and the data question. What should kāwanatanga require of algorithmic instruments of governance, and what does rangatiratanga over taonga mean when the taonga in question is data? Drawing on the work of Te Mana Raraunga, this instalment establishes that any account of public sector governance in Aotearoa that fails to take Te Tiriti seriously is, by definition, incomplete, and misses the unique opportunity to shape the global discourse.

The ethics we already have. The Public Service Commissioner’s reset Code of Conduct, which took effect on 30 March 2026, is neither guidance nor a framework; it is a set of binding minimum standards issued under the Public Service Act 2020. This instalment suggests that the existing instrument is a potential source of algorithmic governance obligations, rather than calling for new principles. The question is whether some existing infrastructure can bear more weight than it is currently asked to.

Aotearoa’s IT record. INCIS, Novopay, and Inland Revenue’s Business Transformation programme tell three very different stories about large-scale IT deployment in Aotearoa. This post asks not simply why the failures failed, but what was different about the institutional conditions in which the IR programme succeeded, and whether those conditions can be generalised.

The friction problem. What does the comparative public administration literature on institutional layering tell us about what happens when non-binding frameworks meet an institutional landscape characterised by marked variation in organisational capability?

Street-level algorithms. What happens to practitioner discretion when algorithmic systems absorb functions that were previously the site of professional judgement? Drawing on the “weaver” construct that Lindsey Te Ata o Tū MacDonald and I have been developing, this post examines how practitioners actively negotiate the boundary between rule and judgement in an increasingly automated environment.

The accountability chain, interrupted. Robodebt, the toeslagenaffaire, the A-level algorithm, Post Office Horizon, the Michigan unemployment fraud system: the cases are accumulating. This post examines what they share, and asks a prior question that the high-profile failures tend to obscure: under what conditions does accountability become performative during adoption, rather than dramatically absent after the fact?

What would adequacy look like? If the series has done its planned work, you will arrive here with a clear picture of what our operating environment demands of algorithmic governance, whether our current arrangements fall short, and, if so, whether we should be worried. This final instalment asks what it would take to close the gap: not more principles, not another framework, but the features under which the current set of principles might actually hold.

What this series is not

It is not a technology commentary. I am not especially interested in what the latest large language model can or cannot do. Nor is it an AI ethics series in the conventional sense, though it engages with ethical questions throughout. It is, at its core, a public administration series: an attempt to bring the tools of institutional analysis, governance, implementation, learning theory, and data sovereignty scholarship to bear on a set of questions that have, thus far, been dominated by technical and legal voices.

It is also not a counsel of despair. The “Pollyanna” label Barb and I used in The Conversation (Te Kawa & Allen, 2026) was a provocation, not a diagnosis, and this series is not built on the assumption that the current arrangements are hopeless. The point is the principles guiding Aotearoa New Zealand’s Public Service AI Framework are important, but principles without institutional and implementation support are, at best, insufficient and, at worst, a mechanism by which central government transfers risk to individual agencies, individual officials, and, most consequentially, the whānau and communities who are subject to algorithmic decisions.

What’s next

I will aim to publish Under the Hood instalments monthly, as a separate series from the other series running on He Poneketanga. Given the pace of this discourse, I will adopt a standing practice of briefly noting significant intervening developments at the top of each instalment where they bear on the argument. The Ontario audit is a case in point: it post-dates the Conversation piece and would have been germane to it.

Each piece will be anchored in specific literature and will aim to include the most current sources available. Suffice to say, this is a fast-moving discourse. Sometimes I will have the latest article; other times I may be a month or so out of date.

The first substantive instalment, on definitional grounds, will follow in a couple of weeks.

In the meantime, the piece that started all of this is available in The Conversation: Pollyanna policy: Is NZ’s framework for AI use in government overly optimistic?

I would be glad to hear what you make of it, and of what you think the series should address. Remember, these are my evolving thoughts, not settled positions. I write this Substack in my spare time. I won’t always get things right. These are think pieces, not position statements. By the end, the series may also provide the groundwork for a more formal paper.

References:

Rahman, S., Connor, J., Dickinson, H., Henne, K., & McDermott, V. (2026). Navigating complexity: A relational perspective on generative AI adoption in government. Australian Journal of Public Administration, 1-14. https://doi.org/10.1111/1467-8500.70044

Office of the Auditor General of Ontario. (2026). Use of artificial intelligence in the Ontario government: Performance audit (Special Report 2026). Toronto: King’s Printer for Ontario.

Te Kawa, D., & Allen, B. (2026, 11 May). Pollyanna policy: is NZ’s framework for AI use in government overly optimistic? The Conversation. https://theconversation.com/polyanna-policy-is-nzs-framework-for-ai-use-in-government-overly-optimistic-281425